gpu farm specs

[theinternet]

according to my calculations, this should crank out around 18k PPD


anything i'm missing?

[theinternet]

with the trim each one of these badboys will run about $1k
you can fill a 42u rack with 10 of em
leave one rack slot for a kvm and one for a switch
youre in for some massive ppd ~180,000/day



only downside is my power bill lmao

[MikeB12]

the gts 250's do about 4k-6k at shader 1890.. you'll rarely see them all on 6k units at the same time.. on average mine did about 5k each when I had all 4 of them on fah.

they do about 9-10k each on gpugrid

also, they have a bad habit of dropping into idle 2d clocks after periodic faulty wu eue's, which do happen occasionally even on a 100% stable card... and the only way to kick it back into full 3d clocks gear is a reboot. so that may be problematic on a big farm with a bunch of gts250's. the 9800gtx+ cards did the same thing. I dont think that issue has ever been resolved. but my 4 have not exhibited that behavior since I moved them to gpugrid, so it has to be something to do with the stanford client and that particular chip (9800/250).. pande wont own up to it though, I've asked on their forums too many times and got the brush off "deal with it" answer everytime.. they keep telling me it's my hardware. and to ask nvidia. I asked nvidia and the only thing they could tell me was to try riva tuner force 3 d clocks. it didn't resolve the behavior and evga would not accept a rma based on that. so go figure, all I can say is based on my experience with fah and gts250 chips..

Unitl fah or nvidia driver comes up with a solution to that "drop into idle mode in while folding" bug, I would consider a different card for folding. they are a nice card for the money, plus you only need one pcie power for each card. but hat little bug is a headache if you like your clients to be left in non-baby-sit mode..

probably a gtx260 216 would be my next choice right now in nvidias lineup, or a gtx275 if you're not concerned about power consumption....

[DR3]

Although you did not mention overclocking the systems, there is still going to be a massive amount of heat to dissipate. You might want to consider an enclosed rack and air management system to force the air through the units and exhaust out the top and into a a/c return duct.
That way the room stays cool and the heat can be better managed.

[Sparky]

Mike already hit it, but I'd recommend a different card than the GTS cards because of the issue he mentioned with them. Although, 70 bucks each is a good price on them (plus the rebate on one of them). Maybe for that price it is worth fiddling with it

[theinternet]

I suppose right now im looking for the best $/ppd ratio for gpu's at the moment. From the reading I've been doing, it seems that the F@H project is moving away from cpu units producing the ppd to gpu units producing the best ppd returns. I realize that the type of work unit farmed out is specialized as to whether it is a cpu or gpu work, but this does seem to be a points game. Most folks I know are migrating their cpu cycles to WCG and building new GPU farms for F@H. I think I'm going to take my cpu's and sling them over to WCG, so I wanted to build a new rack for F@H. The GTX 260's arent much more in terms of $, but building a system with them costs more $ due to 2x pcie plugs and a greater power demand. Let's turn this thread into the best GPU farm that can produce the maximum ppd for the least amount of $.

[theinternet]

this box i plan on building is going to be a 100% full speed ahead dedicated folder, so usability isnt an issue

[shadowwind]

if you do go with the gts250s then go with the 1GB memory cards they got the 55nm chip the cards your looking at have the 65nm chip they are re badge
gtx9800 the 1 GB are the real deal, i owned 6 of them or did i just sold 3 cards last week they was 65nm, probably cheaper to for the gtx 260s in the long run
good luck, and better ppd with the 260s.

[DR3]

Agree with shadowwind to go with the GTS 250 1G models.
Less power and heat than the 512s.

Speaking of power most reviews I have read claim a power consumption of around 148 at idle and 250 at load. Might want to look at a little more power in the PSU. Also the Corsair is a has only a single 12v rail.

Might consider the COOLMAX CUG-950B.
About the same price and has four 12v rails.

Have one myself in my gaming system and it is rock solid.
Also like the semi-modular plugs.

[DR3]

Have you thought about an AMD option?
Possibly the:
ASUS M4N82 Deluxe AM3/AM2+/AM2 NVIDIA nForce 980a SLI AMD Motherboard

with either the:
AMD Athlon II X2 250 Regor 3.0GHz Socket AM3 65W Dual-Core
AMD Phenom II X2 550 Black Edition Callisto 3.1GHz Socket AM3 80W Dual-Core

Would probably run just as good, lower heat and you may even be able to unlock the other two cores for a quad core.

Just an idea.

[DR3]

Just ordered the stuff to make one of these rigs.
Changed a few things, but basically the same thing.



After the rebates it should be under $1K

Will try to get it rolling before I leave for vacation on Thursday.

[123bob]

I suppose right now im looking for the best $/ppd ratio for gpu's at the moment. From the reading I've been doing, it seems that the F@H project is moving away from cpu units producing the ppd to gpu units producing the best ppd returns. I realize that the type of work unit farmed out is specialized as to whether it is a cpu or gpu work, but this does seem to be a points game. Most folks I know are migrating their cpu cycles to WCG and building new GPU farms for F@H. I think I'm going to take my cpu's and sling them over to WCG, so I wanted to build a new rack for F@H. The GTX 260's arent much more in terms of $, but building a system with them costs more $ due to 2x pcie plugs and a greater power demand. Let's turn this thread into the best GPU farm that can produce the maximum ppd for the least amount of $.

And there you go, I'm coming the other way, from WCG to F@H, and run a stable of Q6600s currently. I have been steadily dropping 260-216s in those rigs without ANY problems or disappointments. They are just rock solid, whether for GPUGrid or F@H.

Just keep bringing your CPU cores to WCG and I'll keep dropping 260-216s on F@H.....

Regards,
Bob

[MikeB12]

those 260's are the most troublefree cards I've tried... they are the bulletproof card of the crunching world. low heat and no babysitting. I wish all my cards were 260's.

[DR3]

Well just for the hell of it quickly through it all together on the bench.
Ran fine with I GPU.
So put in the other 2 GPUs.
Got 2 GPUs running and it died when I activated the third.
Now it does not even boot.
Probably overloaded the north bridge or it could be a heating problem.

Will have to troubleshot it after vacation.

[DR3]

Well I got a little time and troubleshot it to a bad power supply.
So hauled out a spare 1200W Toughpower and system now running stable with all 3 GPUs cooking.



Still having a problem with the Nvidia clients not downloading work.
Seems they take forever to get work.
Wonder why the ATI cards do not have this problem?
Do they have different servers for ATI and Nvidia?
If so then they need to get a few more Nvidia ones so they can handle the load.

Well time to send the power supply back to NewEgg or RMA it.

Will be on vacation so will just leave the rigs crunching.

[DR3]

Well it seems to be running. I now have 12 active processors within the last 7 days. This is up 3 from my usual of 9.
Also it looks like I will be breaking my one day ppd mark of 20,987.
Great thing for the team, not so good for Kincaid_Bryan_PC whom I am closing in on.

The server case is huge, but well laid out.
Will work on this once I return.

[AtlasShrugged]

DR3,

Congratulations on the build, looks nice!

You might look at the RAM used by that system, I've found that I can fold 8 GPU clients with just 2GB on XP x64. Take out one of those sticks and you have a head start on your next dedicated GPU folder!

Jason

[DR3]

Got back from vacation yesterday.
Man I was so far in the sticks, I had to go down river 30 miles just to get a cell signal. It was great.

Today I was able to put the Folding rig in the server case.
The case is very roomy and came with 3 nice 80mm fans.

Will try to reduce the RAM when I build the next system.
Also with only one stick of RAM a RAM cooler is not needed.

Have to Return the 1KW power supply and get another one.

[DR3]

Would a 4 card system be more cost effective.

I am thinking it would be the same as the current rig except:

Foxconn Destroyer AM2+/AM2 NVIDIA nForce 780a SLI ATX AMD Motherboard


A Phenom 8650 should work.
Same memory and power supply.

With 4 GTX260s

I have not built an AMD system since the 486 days.
Would this work??

[perkam]

And there you go, I'm coming the other way, from WCG to F@H.

I smell a traitor !!!



If we didn't have people sharing resources between projects, more projects would go the way of D20L and Rosetta

Perkam

[p2501]

Would this work??

You can try, but judging from what AtlasFolder said there seems to be a problems with populating all PCIe slots on that board (apparently a BIOS problem). IF you really want you'd need the MSI board.

But he was trying to run 4 GTX295, maybe single-GPU boards will be different, who knows.

[DR3]

You can try, but judging from what AtlasFolder said there seems to be a problems with populating all PCIe slots on that board (apparently a BIOS problem). IF you really want you'd need the MSI board.

But he was trying to run 4 GTX295, maybe single-GPU boards will be different, who knows.

Are you referring to the MSI 790FX-GD70 AM3?

[p2501]

Right, that was a bit unclear on my part. I was talking about the MSI K9A2 Platinum V2. It's proven to be a stable board for 4 card setups.

[DR3]

I am so confused?
Whats new?

Right, that was a bit unclear on my part. I was talking about the MSI K9A2 Platinum V2. It's proven to be a stable board for 4 card setups.

I am referring to the 790FX-GD70


Are you referring to this motherboard?

I can not find a K9A2 boards with 4 PCIEx16 slots.

[zanzabar]

that gd70 is a nice board, it dosnt unlock the 720 well but it dose the 550